The team from the B2C Digital Assistant tribe at Sber participated in the Kaggriculture competition on Kaggle, where agents manage farms to maximize profit over 30 game days. The team tested six approaches, including rule-based scripts and a transformer model with 20 million parameters, ultimately securing 436th place out of 10,148 teams and earning a silver medal.
The game environment involves a 10x10 field where players manage crops and livestock. Revenue is generated by selling goods to a shared market, where prices fluctuate based on supply and demand. The team developed a simulation environment using JAX to process batches on GPUs and used a ghost-testing setup to compare agent performance against recorded replays from top players.
The team implemented a k-NN agent that mimics the actions of a top player by selecting recorded sequences based on field layout, market prices, and current game state. This approach successfully replicated the early game strategy but struggled in the late game. The team also explored using CP-SAT solvers for seasonal planning, though these models were computationally expensive to recompute every turn.
The final successful approach utilized a transformer architecture inspired by the Orbit Wars competition. The model processes 197 tokens representing game entities, including global features, players, field cells, and market slots. With eight EntityBlocks and a width of 384, the model contains approximately 20.2 million parameters and was trained using behavior cloning.
